Research And Implementation Of Complex Networks Community Detection Methods Based On Graph Neural Network

In the information age,the network is a common structure for representing and analyzing data,such as paper citation networks,Weibo’s user networks and biological networks.In the network,nodes are used to represent entities,and edges are used to represent relationships between entities.After years of research,scholars have found that most networks have the community structure,that is,the network can be divided into a collection of nodes of varying numbers and sizes.The node relationship in the same community is relatively close,and the node attributes are similar;the node relationship between different communities is relatively sparse,and the node attributes are very different.As the complexity of data increases,the network becomes more and more complex.In particular,nodes in the network may have multiple identities and belong to multiple communities.This is called overlapping community networks.With the development of machine learning,neural networks have also been applied to networks,and a new direction called Graph Neural Network was born.An important application of graph neural networks is community detection.However,applying graph neural network to non-overlapping community networks and overlapping community networks still faces many problems.In order to better apply graph neural network to non-overlapping community network community detection and overlapping community network community detection,and to obtain better community division results,this paper completes the following research.And then design and implement graph neural network community detection algorithms for the above two networks.(1)A graph neural network algorithm combining attention mechanism and subgraph strategy is proposed for non-overlapping community network community detection.First,this paper uses the attention mechanism to solve the problem of noise caused by the equal treatment of all neighbors of the node in the feature propagation process of graph neural networks.Then for the shortcomings of graph neural network,especially the graph neural network with attention mechanism added,which has low computational efficiency,subgraph strategy is introduced to significantly reduce the running time of the algorithm without significantly affecting the accuracy of community division.Finally,the attention mechanism and subgraph strategy are combined to implement a non-overlapping community network community detection algorithm.(2)A graph neural network algorithm based on Ego segmentation is proposed for community detection of overlapping community networks.With the help of the idea of Ego splitting,one or more copies of nodes are used to replace the original nodes,and the overlapping community network is transformed into a non-overlapping community network,which effectively simplifies complex problems and separates the identities of the nodes.Then use the unsupervised learning graph neural network algorithm to generate the node embedding of the non-overlapping community network.Finally,according to the mapping between the nodes,feature integration and multilabel classification are performed to obtain the final overlapping community division.(3)Respectively evaluate the performance of the two algorithms in real-world networks.Compared with various community detection algorithms with good performance to verify the accuracy of the community division and the algorithm efficiency of the two algorithms proposed in this paper on their respective community detection problems.The experimental results of real-world networks show that the graph neural network algorithm,which combines the attention mechanism and subgraph strategy designed and implemented in this paper,has achieved excellent results on the problem of non-overlapping community network community detection,and at the same time has greatly improved the efficiency of the algorithm.The graph neural network algorithm based on Ego segmentation designed and implemented in this paper has also achieved excellent results compared with other algorithms on overlapping community network community detection.